Cloud Giants Post Record Q2 Revenue and Backlog; Bernstein Bullish on Amazon and Microsoft
nashnova research
The four major cloud providers posted combined Q2 revenue of $105 billion and a record $2.33 trillion in remaining performance obligations, as Bernstein argues the bottleneck has shifted from GPUs to physical data-center capacity and reiterates Outperform on Amazon and Microsoft.
$105 billion in revenue, $2.33 trillion in backlog — how big is this?
AWS, Azure, Google Cloud and Oracle OCI posted combined cloud revenue of $105 billion in Q2, up 52% year-on-year. Combined remaining performance obligations (RPO — revenue already contracted but not yet recognized) hit $2.33 trillion, up 188%.
This means → customers are not just spending — they are locking in cloud capacity years ahead. Demand is far outpacing supply.
Bernstein estimates the hyperscale cloud market at $1.3–1.5 trillion, making it the largest tangible market in all of software and internet.
The bottleneck has shifted — where is it now?
Bernstein flags that the cloud supply constraint has moved from GPU chip shortages to power-ready physical data-center capacity. In plain terms = the chips are available; building and electrifying the buildings is what can't keep up.
Combined single-quarter capex for Microsoft, Amazon, Google and Oracle reached roughly $150 billion. Bernstein models $660 billion for full-year 2026, approaching $800 billion with Meta included.
On the financing side, every major US cloud company except Microsoft has turned to debt issuance or equity offerings this year. This reflects a building pace so intense that even strong internal cash flow cannot fund it alone.
Does the payback math actually work?
Bernstein cites Amazon's investment framework: data-center shells are built roughly two years ahead and generate revenue for 30+ years. Servers are procured months before deployment, pay back in two to three years on contracts spanning five to six years. Successive hardware generations reuse the same buildings, improving economics each cycle.
AWS's average contract term has risen to 6.4 years. Most 2027 capacity is already locked up, with orders extending into 2028. This means → revenue contracts are signed before the money goes out the door — the risk profile is lower than the market assumes.
Depreciation policies are diverging: Amazon shortened some server useful lives from 6 to 5 years (more conservative), while Microsoft extended data-center useful lives from 15 to 25 years — lowering reported 2026 capex from $190 billion to $175 billion, though the company stresses actual hardware spend is unchanged. In plain terms = same dollars spent, different accounting clocks — Microsoft spreads the cost over longer; Amazon, shorter.
Will AI crush margins?
Bernstein warns that AI will weigh on IaaS and PaaS (basic cloud infrastructure) gross margins at least during the build-out phase. The key variable is the true useful life of GPUs.
Microsoft management recently stated that AI business margins will converge toward non-AI margins over time.
Oracle is passing component price increases — memory and others — through to customers via contract terms, and disclosed that lease renewal prices on three-to-four-year-old GPUs are roughly 20% higher than the original rates. This reflects a market where old equipment is appreciating, not depreciating, because capacity is so tight.
Who is favored and who is not?
Amazon: Outperform maintained, price target $320. AWS grew 37% in Q2 — the fastest in 18 quarters. AI revenue annualized at $25 billion (up from $15 billion last quarter). Management outlined a path toward $1 trillion in annual AWS revenue.
Microsoft: Outperform maintained, price target $660. Key differentiator: guidance for positive free cash flow in fiscal 2027 with no new debt. Oracle: Outperform maintained, price target $325; OCI revenue $7.4 billion (up 121% YoY), RPO $664 billion (roughly $300 billion from OpenAI alone).
Google: the only US hyperscaler not rated Outperform. Bernstein keeps it at Market Perform, target $385 — citing delays to Gemini 3.5 Pro and reliance on leased third-party infrastructure due to capacity constraints, both pressuring margins.
Why is CoreWeave rated Underperform?
CoreWeave posted Q2 revenue of $2.58 billion (up 112% YoY), backlog of roughly $104 billion (up 246%), and raised prices across the board by about 25% in July — strong numbers on the surface.
Yet Bernstein rates it Underperform with a $74 price target. The reasoning: high customer concentration and large clients' stated intent to build their own data centers. In plain terms = CoreWeave's business is a bridge — it fills the gap while hyperscalers can't build fast enough. Once data-center capacity catches up, CoreWeave faces the first and hardest hit.
The key validation checkpoint: Oracle's peak cash requirement is projected for fiscal 2028. Whether revenue growth can cover capex pressure at that point will test the entire bullish thesis.
市场有风险，内容仅供研究参考，不构成投资建议。